基于内容的图像检索算法研究
摘要
基于内容的图像检索(CBIR:Content-Based Image Retrieval)是当前多媒体信息检索的热点之一。提取哪些特征,如何提取特征以进行高效、准确的检索是基于内容的图像检索技术中的核心问题。本文首先介绍了基于内容图像检索技术的理论及其应用研究的整个发展过程,讨论和分析了在图像检索,尤其是基于内容的图像检索领域的技术和现状,并对图像检索系统中的关键技术进行了重点论述。
本文主要对基于内容的图像检索中基于低层特征,包括颜色特征、纹理特征、形状特征的相关算法进行了深入研究,提出了一种新的基于特征进行检索的图像检索算法,算法中颜色特征的提取使用了基于HSV 颜色空间的颜色直方图,纹理特征的提取使用了基于灰度共生矩阵的反差、能量、熵三个纹理特征量,形状特征的提取使用了基于Hu 矩的形状不变矩的一阶矩和二阶矩两个特征量,并使用高斯归一化特征向量。同时,算法中加入了基于HSV 颜色空间的皮肤检测模型,实现了的不良图像检测的功能。
在此基础上,开发实现了具有不良图像检测能力的基于内容的图像检索系统,通过实验测试,得到了较好的检索效果,对于今后的研究具有一定的参考意义和实用价值。
With the development of the multimedia network technology, the application of the image is extensive,the rapid increase of image application, Content-based Image Retrieval (CBIR) becomes most active one in multimediaretrieval field. In order to analysis the information included in an image, the CBIR system always analyses the color,texture, shape, and other low-layer image features, to establish retrieval vectors as retrieval index. In present time, the main CBIR method is similarity retrieval based on multi-dimension feature vector of image. Extracting features from image is the key issues in CBIR.
In this paper, on the base of widely referring to the material about Content-Based Image Retrieval of home and abroad, the whole evolution of the theory on the Content-Based Image Retrieval technique and application study are reviewed at first, then the technology of Content-Based Image Retrieval and current situation are explained in an all-round way, and the key technology in the CBIR system is probed.
In this dissertation I have further studied on the algorithm about the retrieval based on the low-layer image features of CBIR. Then I have present to use the associations of six pieces of features vector that are extract from the three image natural features which are color , texture and shape for form a new retrieval algorithm of using multiple features. The algorithm have been used on the discernment of adult image actually, and have certain discernment results.
The color feature is the most simple and most convenient , effective feature of describing a image. It is the most extensive that use. Based on the human's psychology, the HSV color space is more ocular and easier to accept than the RGB color space. So in this paper the extracting of the color feature have dose in the HSV color space.
While extracting,the image is transformed from the RGB color space to the HSV color space at first. At the same time, the vectors of s and v are mapping to the integer field for calculating amount that sparingly and raising the efficiency. Then the algorithm is calculating the Color Histogram based on the HSV color space and doing the non-equidistant quantization to reduce the dimension. Finally, the vector is the color feature vector which is a part of the multiple features vector.
The texture feature is one of the important feature of the image, usually define a local property of the image, or to a kind of tolerance of the relation between pels in part area. Its essence is a space distribution law of grey level of neighborhood which portrays pels. The grey level co-occurrence matrix is a statistics measurement with two steps of the grey level of the image, and it is a means of the most frequently used to exacting texture features. There are 14 statistical values have been used to exacting texture features in the grey level co-occurrence matrix now. In this dissertation, I only use three of them to improve the calculate efficiency of the algorithm: contrast, energy and entropy. While exacting, the image needs to be transform from a color image to a grey image at first. The method of exacting texture features which based on the grey level co-occurrence matrix is based on the grey values of the image. In this dissertation the images which be used are all the color bitmap
images and it isn’t have a grey value.So we needs to transform the image from the RGB color space to the YUV color space, and the matrix values of Y are the grey values. Then construct two directions co-occurrence matrixs which are θ=0o and θ=90o of the image, calculate the three values have been discussed above, finally the values are the texture feature vectors which are part of the multiple features vector.
The shape feature is the most essential feature to portray objects, and it has to find some geometry invariants mainly for exacting shape features. The moment feature is one of the most important shape features of the image. It’s construct on the statistic analysis of the distribution of grey values which inside a region. It’s a description of the statistic average and the whole feature which can describe objects from the global point. Hu invariant moment is one of the most important shape invariant moments. In the algorithm, because of the shape feature is just the complementarity of the color feature and texture feature and to improve the calculate efficiency, it only used one step moment and two steps moment of the sevev shap invariant moment to be the shape feature vectors which are part of the multiple features vector.
In this paper, the multiple features vector of the retrieval algorithm is a six dimensions vector. Every element of the vector has different magnitude from each other, so it must be a large deflection when measure the similar by the Euclidean distance. The deflection must be eliminate by the vector normalization method. When exacting is over, the algorithm uses the Gaussian normalization method to mormalizied the feature vector. Finally, it uses the Euclidean distance which is simplest and most frequently to use to be
本文主要对基于内容的图像检索中基于低层特征,包括颜色特征、纹理特征、形状特征的相关算法进行了深入研究,提出了一种新的基于特征进行检索的图像检索算法,算法中颜色特征的提取使用了基于HSV 颜色空间的颜色直方图,纹理特征的提取使用了基于灰度共生矩阵的反差、能量、熵三个纹理特征量,形状特征的提取使用了基于Hu 矩的形状不变矩的一阶矩和二阶矩两个特征量,并使用高斯归一化特征向量。同时,算法中加入了基于HSV 颜色空间的皮肤检测模型,实现了的不良图像检测的功能。
在此基础上,开发实现了具有不良图像检测能力的基于内容的图像检索系统,通过实验测试,得到了较好的检索效果,对于今后的研究具有一定的参考意义和实用价值。
With the development of the multimedia network technology, the application of the image is extensive,the rapid increase of image application, Content-based Image Retrieval (CBIR) becomes most active one in multimediaretrieval field. In order to analysis the information included in an image, the CBIR system always analyses the color,texture, shape, and other low-layer image features, to establish retrieval vectors as retrieval index. In present time, the main CBIR method is similarity retrieval based on multi-dimension feature vector of image. Extracting features from image is the key issues in CBIR.
In this paper, on the base of widely referring to the material about Content-Based Image Retrieval of home and abroad, the whole evolution of the theory on the Content-Based Image Retrieval technique and application study are reviewed at first, then the technology of Content-Based Image Retrieval and current situation are explained in an all-round way, and the key technology in the CBIR system is probed.
In this dissertation I have further studied on the algorithm about the retrieval based on the low-layer image features of CBIR. Then I have present to use the associations of six pieces of features vector that are extract from the three image natural features which are color , texture and shape for form a new retrieval algorithm of using multiple features. The algorithm have been used on the discernment of adult image actually, and have certain discernment results.
The color feature is the most simple and most convenient , effective feature of describing a image. It is the most extensive that use. Based on the human's psychology, the HSV color space is more ocular and easier to accept than the RGB color space. So in this paper the extracting of the color feature have dose in the HSV color space.
While extracting,the image is transformed from the RGB color space to the HSV color space at first. At the same time, the vectors of s and v are mapping to the integer field for calculating amount that sparingly and raising the efficiency. Then the algorithm is calculating the Color Histogram based on the HSV color space and doing the non-equidistant quantization to reduce the dimension. Finally, the vector is the color feature vector which is a part of the multiple features vector.
The texture feature is one of the important feature of the image, usually define a local property of the image, or to a kind of tolerance of the relation between pels in part area. Its essence is a space distribution law of grey level of neighborhood which portrays pels. The grey level co-occurrence matrix is a statistics measurement with two steps of the grey level of the image, and it is a means of the most frequently used to exacting texture features. There are 14 statistical values have been used to exacting texture features in the grey level co-occurrence matrix now. In this dissertation, I only use three of them to improve the calculate efficiency of the algorithm: contrast, energy and entropy. While exacting, the image needs to be transform from a color image to a grey image at first. The method of exacting texture features which based on the grey level co-occurrence matrix is based on the grey values of the image. In this dissertation the images which be used are all the color bitmap
images and it isn’t have a grey value.So we needs to transform the image from the RGB color space to the YUV color space, and the matrix values of Y are the grey values. Then construct two directions co-occurrence matrixs which are θ=0o and θ=90o of the image, calculate the three values have been discussed above, finally the values are the texture feature vectors which are part of the multiple features vector.
The shape feature is the most essential feature to portray objects, and it has to find some geometry invariants mainly for exacting shape features. The moment feature is one of the most important shape features of the image. It’s construct on the statistic analysis of the distribution of grey values which inside a region. It’s a description of the statistic average and the whole feature which can describe objects from the global point. Hu invariant moment is one of the most important shape invariant moments. In the algorithm, because of the shape feature is just the complementarity of the color feature and texture feature and to improve the calculate efficiency, it only used one step moment and two steps moment of the sevev shap invariant moment to be the shape feature vectors which are part of the multiple features vector.
In this paper, the multiple features vector of the retrieval algorithm is a six dimensions vector. Every element of the vector has different magnitude from each other, so it must be a large deflection when measure the similar by the Euclidean distance. The deflection must be eliminate by the vector normalization method. When exacting is over, the algorithm uses the Gaussian normalization method to mormalizied the feature vector. Finally, it uses the Euclidean distance which is simplest and most frequently to use to be
引文
[1] Aki.Kobayashi, Toshiyuki Yoshida, Sakai. Image Retrieval by Estimating Parameters. of Distance Measure. SPIE. San Jose, California, 2000, 3972:432-441.
[2] M J Swain, D H Ballard. Indexing via Color Histograms. Proc ICCV90:390-393.
[3] M J Swain, D H Ballard. Color Indexing. IJCV, 1991, 7(1):11-32.
[4] G Pass, R Zabih, J Miller. Comparing Images Using Color Coherence Vectors. Proc of ACM Intern Conf Multimedia. Boston , MA, 1996.
[5] R Rickman, S John. Content Based Image Retrieval Using Color Tuple Histograms. Proc of SPIE:Storage and Retrieval for Image and Video Database. San Jose, CA , 1996, 2670.
[6] J.Huang, S.Kumar, M.Mitra, W.J.Zhu, and R.Zabih. Image Indexing Using Color Correlogram. Proc.of IEEE Conf. On Computer Vision and Pattern Recognition, 1997.
[7] R M Haralick, Shangmugam, Dinstein. Textural Feature for Image Classification. IEEE Trans on Systems. Man , Cybernetics ,1973, SMC-3(6):610-621.
[8] Hideyuki Tamura, et al. Texture Features Corresponding to Visual Perception[J]. IEEE Trans, on Systems, Man and Cybemetics, 1978, 8(6):460-473.
[9] R M Haralick. Statistical and Structural Approaches to Texture[J]. Proc.IEEE, 1979, 67:786-804.
[10] Jianchang Mao, Anil K Jain. Texture Classification and Segmentation Using Multiresolution Simultaneous Autoregressive Models[J]. Pattern Recognition, 1992, 25(2):173-188.
[11] 郑南宁. 计算机视觉与模式识别[M]. 北京:国防工业出版社,1998:90-92.
[12] B S Manjunath, W Y Ma. Texture Features for Browsing Retrieval of Image Data[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1996, 18(8):837-842.
[13] Michael Diaz. Wavelet Features for Color Image Classification[A]. Imaging and Geospatial Information Society[C]. 2000 Annual Conference, Orlando, 2000.
[14] S Bres, R Schettini. Detection of Interest Points for Image Indexation[A]. Third International Conference, VISUAL’99[C]. Amsterdam, the Netherlands, 1999:427-434.
[15] Y Tao, W I Grosky. Delaunay Triangular for Image Object Indexing: Anovel Method for Shape Representation[A]. SPIE Conference on Storage and Retrieval for Image and Video Database[C]. San Jose,1999:631-642.
[16] A K Jain, A Vailaya. Image Retrieval Using Color and Shape[J]. Pattern Recognition, 1996, 29(8):1233-1244.
[17] A K Jain, Yu Zhong, Scridhar Lakshmanan. Object Matching Using Deformable Templates[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1996, 18(3):267-278.
[18] A Del Bimbo, et al. Image Retrieval by Elastic Matching of Shapes and Image Patterns[A]. ICMCS[C]. Japan: Hiroshima, 1996:215-218.
[19] T Kohonen. Automatic Formation of Topologically Correct Feature Maps. Biol, Cybern , 1982, 43:59-69.
[20] J J Hopfield. Neurons with Graded Response have Collective Computational properties like those of two state Neurons. Proc. Natl, Acad. Sci., 1984:3088-3092.
[21] D F Specht. Probabilistic neural networks and the polynomial Adaline as complementary techniques for classification. IEEE Trans on Neural Networks, 1990, 1(1):111-121.
[22] Y Rui, T S Huang, M Ortega, et al. Relevance feedback: A power tool for interactive content-based image retrieval. IEEE Trans on Circuits and Video Technology, 1998, 8(5).
[23] M K Mandal, F Idris, S Panchanathan. A Critical Ecaluation of Image and Video Indexing Techniques in the Compressed Domain[J]. Image and Vision Computing, 1999, 17(7):513-529.
[24] M Shneier,M Abdel-Mottaleb. Exploiting the JPEG Compression Scheme for Image Retrieval[J]. IEEE Trans. on Parrern Analysis and Machine Intelligence, 1996, 18(8):849-853.
[25] J Jiang, A Armstrong, G C Feng. Direct Content Access and Extraction from JPEG Compressed Images[J]. Pattern Recognition, 2002, 35:2511-2519.
[26] Charles E Jacobs, et al. Fast Multiresolution Image Querying[A]. Proceedings of SIGGRAPH’95[C]. Los Angels,CA,1995:277-286.
[27] Guojun Lu, Shyhwei Teng. A Novel Image Retrieval Technique Based on Vector Quantization[A]. Proceedings of International Conference on Computational Intelligence for Modelling, Control and Automation[C].
[2] M J Swain, D H Ballard. Indexing via Color Histograms. Proc ICCV90:390-393.
[3] M J Swain, D H Ballard. Color Indexing. IJCV, 1991, 7(1):11-32.
[4] G Pass, R Zabih, J Miller. Comparing Images Using Color Coherence Vectors. Proc of ACM Intern Conf Multimedia. Boston , MA, 1996.
[5] R Rickman, S John. Content Based Image Retrieval Using Color Tuple Histograms. Proc of SPIE:Storage and Retrieval for Image and Video Database. San Jose, CA , 1996, 2670.
[6] J.Huang, S.Kumar, M.Mitra, W.J.Zhu, and R.Zabih. Image Indexing Using Color Correlogram. Proc.of IEEE Conf. On Computer Vision and Pattern Recognition, 1997.
[7] R M Haralick, Shangmugam, Dinstein. Textural Feature for Image Classification. IEEE Trans on Systems. Man , Cybernetics ,1973, SMC-3(6):610-621.
[8] Hideyuki Tamura, et al. Texture Features Corresponding to Visual Perception[J]. IEEE Trans, on Systems, Man and Cybemetics, 1978, 8(6):460-473.
[9] R M Haralick. Statistical and Structural Approaches to Texture[J]. Proc.IEEE, 1979, 67:786-804.
[10] Jianchang Mao, Anil K Jain. Texture Classification and Segmentation Using Multiresolution Simultaneous Autoregressive Models[J]. Pattern Recognition, 1992, 25(2):173-188.
[11] 郑南宁. 计算机视觉与模式识别[M]. 北京:国防工业出版社,1998:90-92.
[12] B S Manjunath, W Y Ma. Texture Features for Browsing Retrieval of Image Data[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1996, 18(8):837-842.
[13] Michael Diaz. Wavelet Features for Color Image Classification[A]. Imaging and Geospatial Information Society[C]. 2000 Annual Conference, Orlando, 2000.
[14] S Bres, R Schettini. Detection of Interest Points for Image Indexation[A]. Third International Conference, VISUAL’99[C]. Amsterdam, the Netherlands, 1999:427-434.
[15] Y Tao, W I Grosky. Delaunay Triangular for Image Object Indexing: Anovel Method for Shape Representation[A]. SPIE Conference on Storage and Retrieval for Image and Video Database[C]. San Jose,1999:631-642.
[16] A K Jain, A Vailaya. Image Retrieval Using Color and Shape[J]. Pattern Recognition, 1996, 29(8):1233-1244.
[17] A K Jain, Yu Zhong, Scridhar Lakshmanan. Object Matching Using Deformable Templates[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1996, 18(3):267-278.
[18] A Del Bimbo, et al. Image Retrieval by Elastic Matching of Shapes and Image Patterns[A]. ICMCS[C]. Japan: Hiroshima, 1996:215-218.
[19] T Kohonen. Automatic Formation of Topologically Correct Feature Maps. Biol, Cybern , 1982, 43:59-69.
[20] J J Hopfield. Neurons with Graded Response have Collective Computational properties like those of two state Neurons. Proc. Natl, Acad. Sci., 1984:3088-3092.
[21] D F Specht. Probabilistic neural networks and the polynomial Adaline as complementary techniques for classification. IEEE Trans on Neural Networks, 1990, 1(1):111-121.
[22] Y Rui, T S Huang, M Ortega, et al. Relevance feedback: A power tool for interactive content-based image retrieval. IEEE Trans on Circuits and Video Technology, 1998, 8(5).
[23] M K Mandal, F Idris, S Panchanathan. A Critical Ecaluation of Image and Video Indexing Techniques in the Compressed Domain[J]. Image and Vision Computing, 1999, 17(7):513-529.
[24] M Shneier,M Abdel-Mottaleb. Exploiting the JPEG Compression Scheme for Image Retrieval[J]. IEEE Trans. on Parrern Analysis and Machine Intelligence, 1996, 18(8):849-853.
[25] J Jiang, A Armstrong, G C Feng. Direct Content Access and Extraction from JPEG Compressed Images[J]. Pattern Recognition, 2002, 35:2511-2519.
[26] Charles E Jacobs, et al. Fast Multiresolution Image Querying[A]. Proceedings of SIGGRAPH’95[C]. Los Angels,CA,1995:277-286.
[27] Guojun Lu, Shyhwei Teng. A Novel Image Retrieval Technique Based on Vector Quantization[A]. Proceedings of International Conference on Computational Intelligence for Modelling, Control and Automation[C].